Let $P(z)$ be a polynomial of degree $n$ with $|P(z)|\leq 1$ on $|z|=1$ and $P_m(z)$ be a partial sum of $P(z).$ How large $P_m(z)$ can be on $|z|=1?$

  • 1
    $\begingroup$ Trivial upper bound is $\|1+e^{i \theta}+...+e^{im \theta}\|_{L^{1}(\mathbb{T})}$. $\endgroup$ May 25 '20 at 4:43
  • 4
    $\begingroup$ What here is fixed and what is allowed to vary? $\endgroup$
    – user44191
    May 25 '20 at 4:46

The trivial upper bound $\max_{|w|=1}|P_{m}(w)|\leq \|1+z+\cdots+z^{m}\|_{L^{1}(\mathbb{T})} \asymp C \log(m)$ that I wrote in the comment is actually sharp in the regime $m=n/2$. Here is the proof.

Notice that $P_{m}(z)$ is convolution of $P(z)$ with $D_{m}(z) = 1+z+\cdots+z^{m}$ on the unit circle, therefore, by the triangle inequality $\max_{|z|=1}|P_{m}(z)| \leq \frac{1}{2\pi}\int_{-\pi}^{\pi}|1+e^{i\theta}+\cdots+e^{i m \theta}| d\theta \asymp C \log(m)$

Next, let us show that the upper bound is sharp in the regime $m=\frac{n}{2}$ where $n$ is large.

Indeed, consider the polynomial $$ P(z) = z^{n} \overline{\left( 1+\frac{z}{1}+\cdots+\frac{z^{n}}{n}\right)} - z^{n}\left(1+\frac{z}{1}+\cdots+\frac{z^{n}}{n} \right) = z^{n}+\frac{z^{n-1}}{1}+\cdots+\frac{1}{n}-z^{n}-\frac{z^{n+1}}{1}-\cdots-\frac{z^{2n}}{n} $$ it is of degree $2n$, and $\max_{|z|=1}|P_{n-1}(z)|\geq |P_{n-1}(1)| \asymp \log(n)$. On the other hand let us show that $\max_{|z|=1}|P(z)|\asymp 1$. Indeed, for $z=e^{ix}$ we have
$$ |P(z)| = 2\left| \sum_{k=1}^{n} \frac{\sin(kx)}{k}\right|\leq 2 \int_{0}^{\pi} \frac{\sin(s)}{s}ds \asymp 1 \quad \text{for all} \quad n \geq 1, \; x \in [0, 2\pi) $$ The last inequality I guess is known "since Democritus".

  • 1
    $\begingroup$ @user159888 I do not see typos. One small thing is that instead of P_{n-1} I should be writing P_{n}. But both are fine it just gives example of degree 2n with m=n-1 $\endgroup$ May 26 '20 at 12:43
  • 1
    $\begingroup$ I haven't looked closely, but it feels like this argument might symmetrize to get an upper bound of $O(\log(\min(m,n-m)))$? $\endgroup$ May 26 '20 at 15:17
  • 1
    $\begingroup$ Yes, this can be done by writing $P_{m}(z) = P(z)-(P(z)-P_{m}(z))$ the first term is estimated by $1$, and the term $(P(z)-P_{m}(z))$ is upper bounded by $\|z^{m+1}+...+z^{n}\|_{L^{1}(\mathbb{T})}\asymp \log(n-m)$. $\endgroup$ May 27 '20 at 2:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.